HugeTLBfsの実用性向上と並列処理への適用と評価

概要

論文の詳細を見る
HugeTLBfs improves performance of the applications which need huge data area by using Huge Page because of reducing TLB misses that affect an efficiency of address translation. However, Huge Page sizes which can be used concurrently are limited to only one size. If a pretty large size in Huge Pages is adopted to reduce TLB misses dramatically, its size is allocated even to the remaining fragment in the last part of data area. As a result of allocating a large size to small area, large useless data area that any processes can not use is generated. On the other hand, if a small size is adopted, improvement of performance is trivial. Then we added the second Huge Page size to fill fragment that is concurrently available with the large ones. Consequently, our method makes it possible to minimize useless data area without going down in performance.
社団法人電気学会の論文